Examples are given here: https://cran.r-project.org/web/packages/sandwich/vignettes/sandwich-OOP.pdf

]]>When I was first thought about the use of 1/var weights it stroke me as a sort of technique of “data dredging”, i.e. improving the model by adding parameters (weights) for which we have no explanation whatsoever. I guess its OK if you only care about prediction (machine learning-like) and not so much about getting a handle on/estimating causal effects… I also recall having the same kind of doubts in the case of Ridge regression.

Anyway, meta-analysis seems to be a somewhat different context.

]]>Cheers!

]]>In my field (reproductive science), you see at least 10 new publications per month where the data is extremely heteroscedastic (mostly increasing variance with increasing magnitude) but linear regressions are based on unweighted fits…

Cheers,

Andrej